{"id":"35264360-cba8-45f6-96ac-ad0237ef09ff","arxiv_id":"1908.01596","paper_version":5,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Doubly stochastic graph shifts are shown to be bounded in expectation and to converge to the mean for i.i.d. signals as neighborhoods grow, though the 'isometry' label is an overstatement.","lead":"The paper analyzes doubly stochastic matrices as graph shift operators and proves bounds on the expected energy of shifted random graph signals. A smart generalist should read it because stable shift operators are a prerequisite for filtering and stationarity analysis in graph signal processing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central i.i.d. consistency claim is false: a doubly stochastic S with S_11=1/2 and S_1n=1/(2(N-1)) gives var{S(x_1)}->sigma^2/4, not 0; Remark 6's fixed L,U also conflicts with N_m L<=1.","rationale":"Read in good faith: the finite-N algebra leading to (15) and the mean-preservation property are correct, and the Sinkhorn fixed-point observation gives independent support that the constructed counterexample lies inside the class. The load-bearing flaw is the asymptotic step. Eq. (17)'s uniform lower bound L>0 cannot coexist with N_m->infinity because each row sums to 1. Therefore the limit in (19) is not a consequence of the stated assumptions. The explicit doubly stochastic matrix above shows the claimed i.i.d. consistency is actually false without an extra balance condition such as max_n S_mn->0; with such a condition the result becomes a variance-annihilation statement, not an L2-isometry. The single-run numerical section does not test the asymptotic regime and cannot repair the theoretical claim. Because the central advertised property (ii) is false as stated and the remaining isometry claim is mislabeled, the conditional acceptance should be revised to reject. This is a technical assessment of the mathematics, not a judgment on the authors.","tokens_in":8616,"tokens_out":16120,"duration_ms":176309,"concrete_test":"Analytic test: set W_N equal to the symmetric matrix S_N with S_{1,1}=1/2, S_{1,n}=S_{n,1}=1/(2(N-1)) for n=2,...,N, and S_{j,n}=(1-1/(2(N-1)))/(N-1) for j,n>=2. Since S_N is positive and doubly stochastic, Algorithm 1 returns S_N unchanged. Compute for i.i.d. x_n~N(mu,sigma^2) the exact variance var{S(x_1)}=sigma^2(1/4+1/(4(N-1))). Let N->infinity; if the limit is sigma^2/4 rather than 0, Remark 7 and Eq. (20) are false. A numerical companion: implement Alg. 1 on W_N for N=2,4,...,1024 and record sum_n S_{1n}^2; it should approach 1/4, not 0.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's asymptotic theorem rests on Remark 6, which asserts that L and U in (17) are invariant to N_m. This is not merely unproven; it is incompatible with N_m->infinity. For row m, (17) gives S_mn >= L > 0 for all N_m entries, so 1 = sum_{n in V_m} S_mn >= N_m L, forcing N_m <= 1/L. Thus any sequence with N_m->infinity must have L_N->0, and the Kantorovich constant K=(L+U)^2/(4LU) is N-dependent. The replacement of K by a fixed constant in the passage from (18) to (19) is therefore unjustified. The failure is not hypothetical: for the symmetric doubly stochastic matrix S with S_11=1/2, S_1n=S_n1=1/(2(N-1)) for n>1, and S_jn=(1-1/(2(N-1)))/(N-1) for j,n>=2, row 1 has sum of squares equal to 1/4 + 1/(4(N-1)). For i.i.d. vertex signals with variance sigma^2 this makes var{S(x_1)}->sigma^2/4, contradicting Remark 7 and Eq. (20). Hence the advertised asymptotic consistency for i.i.d. signals does not hold for the stated class of doubly stochastic GSOs. In addition, the 'L2-isometry' label is wrong: even when E{S(x_m)^2}->mu^2, the unshifted signal has E{x_m^2}=mu^2+sigma^2, so the operator is norm-contracting, not isometric.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a class of doubly stochastic graph shift operators (GSOs) and claims three properties: (i) lower and upper L2-boundedness for locally stationary random graph signals, (ii) L2-isometry for i.i.d. random graph signals in the asymptotic limit of growing incoming neighbourhoods, and (iii) preservation of the graph signal mean. The theoretical development models the shifted vertex signal as a weighted combination of neighbouring vertex random variables, derives an upper bound on the shifted variance via Cauchy-Schwarz and Kantorovich inequalities, and then takes the limit N_m→∞ to obtain variance vanishing for ρ=0 and a finite bound for ρ>0. A numerical example on temperature sensor data is used to illustrate the denoising effect of the proposed shift.","tokens_in":9019,"tokens_out":6399,"duration_ms":64244,"significance":"If the asymptotic consistency and isometry results were correct, they would provide a useful and simple class of graph shift operators for graph signal processing, with an appealing connection to Markov chains and Sinkhorn-Knopp normalization. The paper also correctly observes that doubly stochastic matrices preserve the mean of any graph signal and preserve the L1 norm of nonnegative signals via the Birkhoff-von Neumann decomposition. However, the central asymptotic claims---variance vanishing for i.i.d. signals and the resulting 'L2-isometry'---are not valid for the stated class of all doubly stochastic matrices. The flaw is load-bearing: it affects Eqs. (19)-(21) and (24), which are the main advertised contributions. Because the error can only be repaired by adding substantive new assumptions and redefining the class of operators, the paper in its current form does not establish its fundamental claims.","major_comments":[{"comment":"The limit in Eq. (19) rests on Remark 6, which asserts that the lower and upper bounds L and U in (17) are invariant to the neighbourhood size N_m. This is not proven and is in fact inconsistent with N_m→∞. For any row m, 1 = ∑_{n∈V_m} S_mn ≥ N_m L, so any sequence of doubly stochastic matrices with N_m→∞ must have L = L(N_m) → 0. Hence the Kantorovich constant K=(L+U)^2/(4LU) depends on N_m, and the replacement of K by a fixed constant in the passage from (18) to (19) is unjustified. Consequently the variance bound (19), the consistency claim (20), and the boundedness results (21) and (24) are not established for the stated class of doubly stochastic GSOs.","section":"III-B, Eqs. (15)-(19), Remark 6"},{"comment":"The claimed asymptotic consistency for i.i.d. signals is false for the stated class. Consider the N×N doubly stochastic matrix S with S_11=1/2, S_1n=S_n1=1/(2(N-1)) for n>1, and S_jn=(1-1/(2(N-1)))/(N-1) for j,n≥2. For i.i.d. vertex signals with variance σ^2, the shifted variance at vertex 1 is σ^2(1/4 + 1/(4(N-1))), which tends to σ^2/4, not 0. This directly contradicts Eq. (20) and Remark 7. Additional hypotheses, such as uniform vanishing of the maximum row entry, are required for the consistency result to hold; they are absent from the paper.","section":"III-B, Remark 7 and Eq. (20)"},{"comment":"The description of Eq. (24) as an 'L2-isometry' is incorrect. The original vertex signal has E{x_m^2}=μ^2+σ^2, whereas the limiting expected power of the shifted signal is μ^2. The operator is therefore norm-contracting in expectation, not norm-preserving; the result describes convergence of S(x_m) to the constant mean μ, a projection, rather than an isometry. This terminology appears in the abstract, introduction, and Remark 9, and materially misrepresents the mathematical content even under additional assumptions that would make the variance vanish.","section":"III-D, Remark 9 and Eq. (24)"}],"minor_comments":[{"comment":"The limit subscript 'N_n→∞' should read 'N_m→∞'; the neighbourhood size of vertex m is the quantity being increased.","section":"III-B, Remark 7"},{"comment":"The statement that ‖S‖_2=1 for every doubly stochastic matrix does not follow merely from the largest eigenvalue being equal to 1; it follows from the Birkhoff-von Neumann decomposition used in Remark 4 (or from the convexity of the spectral norm on permutation matrices). The argument as written is incomplete.","section":"II-C, Remark 2"},{"comment":"The L1-isometry statement should explicitly restrict to nonnegative graph signals; for signed signals the L1 norm of Sx can be strictly smaller than that of x, as the mean-preservation argument in Remark 3 shows.","section":"II-C, Remark 4"},{"comment":"The example does not report the values of N_m, L, or U for the Sinkhorn-Knopp normalized matrix, so the connection between the theoretical bounds and the demonstrated 5.8 dB SNR gain remains purely illustrative rather than quantitative.","section":"IV, Numerical example"}],"recommendation":"reject","confidential_remarks":"The counterexample in major comment 2 is easy to verify and directly disproves the advertised consistency and isometry claims for the class of all doubly stochastic matrices. Because these claims are the paper's central theoretical contribution, a revision would require redefining the operator class and reworking the abstract and conclusions; this goes beyond a standard major revision. The paper would be more suitable as a shorter note on the specific boundedness inequalities that do hold under explicit assumptions on L and U."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the central claim — that doubly stochastic graph shifts become asymptotically consistent and L2-isometric for i.i.d. random graph signals — does not hold for the stated class. The proof hinges on Remark 6, which asserts the lower and upper bounds L and U on shift entries are invariant to the neighborhood size N_m. That is not merely unproven; it is incompatible with row sums. If every entry in a row is at least L > 0, then 1 = sum_n S_mn >= N_m L, so any sequence with N_m -> infinity forces L_N -> 0. The stress-test counterexample lands: for the doubly stochastic S with S_11 = 1/2, S_1n = 1/(2(N-1)) for n>1, and the remaining entries filled to maintain double stochasticity, i.i.d. zero-mean vertex signals give var{S(x_1)} -> sigma^2/4, contradicting eq. (20). So the advertised consistency for i.i.d. signals is false. What the paper does well: it correctly observes that doubly stochastic GSOs preserve the mean and are L_p contractions for p=1,2,infty. The fixed-N variance bound using Cauchy-Schwarz and Kantorovich (eqs. 15-16) is a reasonable way to bound the second moment, and the local stationarity model with common mean, variance, and positive correlation is a useful simplification. The framing of the GSO as an unbiased expectation operator is nice. Soft spots: Eq. (18) claims the Kantorovich bound is <1, but that is not true when (L+U)^2/(4LU) > N_m, which happens for small neighborhoods or heterogeneous weights. The label L2-isometry is wrong: even if E{S(x_m)^2} -> mu^2, the input signal has power mu^2 + sigma^2, so the operator is norm-contracting, not isometric. The numerical example is one run, no error bars or baselines, and it does not approach the asymptotic regime. All of these are addressable, but the central theorem needs repair. Who this is for: graph signal processing readers interested in diffusion-based shift operators. The boundedness observations are correct, but the asymptotic consistency claim is not. Recommendation: deserves a serious referee, not a desk reject, because the flaw is subtle and fixable — restrict to a subclass where row masses concentrate, or reframe as convergence to the mean. But as written I would not accept without major revision.","headline":"The paper's headline asymptotic consistency result for doubly stochastic GSOs is false as stated; fixed-N bounds are fine, but the limit argument rests on an impossible invariance assumption and the L2-isometry label is wrong.","tokens_in":748,"tokens_out":1835,"would_cite":false,"duration_ms":56168,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a class of doubly stochastic graph shift operators and shows that, for locally stationary random graph signals, they are asymptotically $L_2$-bounded, and asymptotically $L_2$-isometric for i.i.d.","keywords":["graph signal processing","doubly stochastic matrix","graph shift operator","statistical consistency","L2 boundedness","L2 isometry","locally stationary graph signals","Kantorovich inequality"],"falsifier":"Take a family of graphs with increasing incoming neighborhood size $N_m$, such as directed stars with edge weights that decay along the leaves, compute the doubly stochastic normalization, and track $\\sum_n S_{mn}^2$ together with the factor $(L+U)^2/(4LU)$; if $N_m\\sum_n S_{mn}^2$ does not stay bounded or $L$ tends to $0$ while $U$ does not, then the claimed limiting bound in eq. (21) fails for that family.","tokens_in":8442,"feed_emoji":"📊","tokens_out":12522,"duration_ms":110677,"temperature":0.7,"pith_summary":"The paper proposes using doubly stochastic matrices---nonnegative matrices whose rows and columns each sum to one---as graph shift operators, and asks what happens to the energy of a random graph signal when such a shift is applied. It establishes three properties: the shift preserves the signal mean exactly; for locally stationary random signals it is upper and lower bounded in expected squared amplitude, with the bound governed by the signal's correlation $\\rho$ and the ratio of the largest to smallest shift entries; and for i.i.d. signals it becomes an $L_2$-isometry in the limit of large incoming neighborhoods, meaning the expected power of the shifted signal converges to the square of the mean. The interest is that ordinary adjacency or Laplacian shifts can distort signal spectra, while this class gives bounded, nearly isometric shifts and comes with a practical averaging and denoising interpretation.","feed_headline":"Doubly stochastic graph shifts preserve signal power in the limit","feed_subtitle":"For i.i.d. signals, expected power converges to the mean squared; correlated signals stay bounded.","key_machinery":"The central object is the doubly stochastic graph shift operator $S\\in\\mathbb{R}^{N\\times N}$, defined by $S_{mn}\\ge 0$, $S\\mathbf{1}=\\mathbf{1}$, $S^T\\mathbf{1}=\\mathbf{1}$. Its left-stochastic half gives a Markov diffusion interpretation, since each column is a set of transition probabilities of a random walker; its right-stochastic half makes each output an unbiased expectation operator, since each row sums to one. The quantitative engine is the Kantorovich inequality applied to the squared row entries, which converts the bounds $L$ and $U$ on shift entries into the factor $(L+U)^2/(4LU)$; the AM-GM inequality then identifies this factor as the squared ratio of the arithmetic to the geometric mean of $L$ and $U$. This machinery reduces boundedness of the shift to a statistical consistency analysis of a graph-dependent average.","core_discovery":"On its own terms, the paper's central discovery is that the doubly stochastic property turns a graph shift into a statistically consistent estimator: each output $S(x_m)=\\sum_{n\\in V_m} S_{mn}x_n$ is an unbiased estimate of the local mean $\\mu$, and its variance can be controlled. Using the Kantorovich inequality, the paper shows that $\\sum_{n\\in V_m} S_{mn}^2 \\le \\frac{1}{N_m}\\frac{(L+U)^2}{4LU}$, with $0